Home
About
Posts
Snippets
Insights
blog logoJooojub
Home
About
Posts
Snippets
Insights

Daily Briefing

April 19, 2026
AI Insightsbriefing
2026-04-18
42 articles

Nvidia's Huang warns DeepSeek running on Huawei chips would be 'horrible' for the US

2026-04-18

Summary

Nvidia CEO Jensen Huang warned that DeepSeek optimizing AI models for Huawei chips would be a 'horrible outcome' for the United States.

Key Points

  • Nvidia CEO Jensen Huang expressed concern about DeepSeek optimizing AI models for Huawei's Ascend chips.
  • DeepSeek is set to release its V4 foundation model on Huawei's Ascend 950PR processor.
  • This move could undermine the software-hardware dependency that underpins US AI dominance.
  • Huang warned that China could gain an advantage over the US through its own AI standards and technology.
  • DeepSeek's models may have been trained on Nvidia chips, but are set to be deployed on Huawei chips.
Notable Quotes & Details

Intended Audience

AI industry stakeholders, investors, policy makers

Anthropic's Amodei meets Wiles and Bessent at the White House in first step toward resolving Mythos standoff

2026-04-18

Summary

Anthropic CEO Dario Amodei held talks with White House officials regarding access to the Mythos AI model.

Key Points

  • Anthropic CEO Dario Amodei met with White House Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent.
  • The meeting was described as a 'productive and constructive' discussion about access to the Mythos AI model.
  • Mythos is a cutting-edge AI model capable of finding thousands of zero-day vulnerabilities.
  • This meeting is the first step toward resolving the situation in which Anthropic was blacklisted by the Department of Defense over safety restrictions.
  • If an agreement is reached, the Department of Defense may be excluded and access to Mythos granted through civilian agencies.
Notable Quotes & Details

Intended Audience

AI industry stakeholders, policy makers, technology and security experts

Palantir, Thales, and a startup are competing to build the FAA's predictive air traffic AI

2026-04-18

Summary

The FAA is competing with Palantir, Thales, and Air Space Intelligence to develop SMART, an AI system for air traffic control.

Key Points

  • The FAA is developing an AI system called SMART (Strategic Management of Airspace Routing Trajectories).
  • SMART aims to extend air traffic collision prediction time from the current 15 minutes to 2 hours.
  • Three companies — Palantir, Thales, and Air Space Intelligence — are competing for the contract.
  • The project is being driven to address the LaGuardia airport collision incident and air traffic controller overwork issues.
  • SMART uses high-precision 4D modeling to predict bottlenecks and schedule conflicts before aircraft take off.
Notable Quotes & Details

Notable Data / Quotes

  • 32.5 billion: FAA modernization program budget
  • 612: number of aging radar systems to be replaced
  • 1,200: number of new controllers to be hired in fiscal year 2026
  • 2026: Palantir's estimated revenue of approximately $7.2 billion

Intended Audience

Aviation industry stakeholders, AI technology developers, government agency officials

Cursor is raising $2 billion at a $50 billion valuation as AI coding tools become the fastest-growing software category

2026-04-18

Summary

AI coding startup Cursor is pursuing $2 billion in funding at a $50 billion valuation, leading the rapid growth of the AI coding tools market.

Key Points

  • AI coding startup Cursor (Anysphere) is in discussions to raise at least $2 billion in funding, co-led by Andreessen Horowitz, Thrive Capital, and Nvidia.
  • The estimated enterprise value is $50 billion, nearly double the $29.3 billion valuation from five months ago.
  • Cursor reached $2 billion in annual recurring revenue (ARR) in just three years, setting the record as the fastest-growing B2B software company in history.
  • It has over 1 million paying customers and over 2 million total users, with approximately 70% of Fortune 1,000 companies as clients.
  • Competition with GitHub Copilot, Claude Code, and others is intensifying.
Notable Quotes & Details

Notable Data / Quotes

  • $2 billion: funds currently under discussion
  • $50 billion: estimated enterprise value
  • $29.3 billion: enterprise value in November 2025
  • 3 years: time to reach $2 billion ARR from zero
  • 1 million+: number of paying customers
  • 70%: percentage of Fortune 1,000 companies as clients
  • August 2024: Series A ($400 million valuation)
  • 5 months later: Series B ($2.6 billion valuation)
  • May 2025: Series C ($9 billion valuation)
  • November 2025: Series D ($29.3 billion valuation)

Intended Audience

Software developers, AI technology investors, startup stakeholders

Three more senior executives leave OpenAI as the company kills its side quests

2026-04-18

Summary

Three senior executives — former CPO Kevin Weil, Sora development lead Bill Peebles, and enterprise CTO Srinivas Narayanan — departed OpenAI alongside the company's discontinuation of 'side quests' and pivot to enterprise AI.

Key Points

  • Three senior executives have left OpenAI.
  • OpenAI is discontinuing 'side quests' such as Sora and OpenAI for Science, and focusing on enterprise AI.
  • This is part of a continuing pattern of leadership attrition, with only 2 of 11 co-founders remaining after two years.
  • Departing executives are moving to Anthropic, Meta's Superintelligence Labs, and various startups.
  • OpenAI is targeting $25 billion in annual revenue while facing projected losses of $14 billion.
Notable Quotes & Details

Notable Data / Quotes

  • $25 billion in annualised revenue
  • projected $14 billion losses
  • 2 of 11 co-founders remain
  • Sora discontinuing 26 April

Intended Audience

AI industry analysts, investors, corporate strategists

Anthropic's relationship with the Trump administration seems to be thawing

2026-04-18

Summary

Despite being designated a supply chain risk by the Department of Defense, Anthropic is showing signs of a thaw in its relationship with senior officials in the Trump administration.

Key Points

  • Anthropic continues to communicate with the Trump administration despite the Department of Defense's supply chain risk designation.
  • Treasury Secretary Scott Bessent and Fed Chair Jerome Powell encouraged major banks to test Anthropic's Mythos model.
  • Anthropic co-founder Jack Clark noted that the supply chain risk designation is a 'narrow contractual dispute' and would not affect government briefings.
  • According to Axios reporting, Bessent and White House Chief of Staff Susie Wiles met with Anthropic CEO Dario Amodei, with the White House describing it as a 'productive and constructive' meeting.
  • The dispute between Anthropic and the Department of Defense began when Anthropic attempted to maintain safeguards against use of its models for fully autonomous weapons and large-scale domestic surveillance.
Notable Quotes & Details

Intended Audience

AI policy analysts, government officials, AI company leaders

Google AI Releases Auto-Diagnose: An Large Language Model LLM-Based System to Diagnose Integration Test Failures at Scale

2026-04-18

Summary

Google AI released Auto-Diagnose, an LLM-based system for diagnosing large-scale integration test failures, helping developers find the root causes of bugs in log files.

Key Points

  • Google AI introduced Auto-Diagnose, an LLM-based tool for diagnosing integration test failures.
  • The tool automatically reads failure logs, finds root causes, and posts concise diagnoses in code reviews.
  • In manual evaluations, it identified root causes with 90.14% accuracy across 71 real-world failure cases.
  • In a Google developer survey, diagnosing integration test failures was one of the top five complaints.
  • 38.4% of respondents said diagnosing integration test failures takes more than an hour, and 8.9% said it takes more than a day.
Notable Quotes & Details

Notable Data / Quotes

  • 90.14% (accuracy)
  • 71 real-world failures
  • 39 distinct teams
  • 52,635 distinct failing tests
  • 224,782 executions
  • 91,130 code changes
  • 22,962 distinct developers
  • 5.8% ('Not helpful' rate)
  • 78% of integration tests at Google are functional
  • 38.4% of integration test failures take more than an hour to diagnose
  • 8.9% take more than a day

Intended Audience

Software developers, QA engineers, DevOps engineers

A End-to-End Coding Guide to Running OpenAI GPT-OSS Open-Weight Models with Advanced Inference Workflows

2026-04-18

Summary

This tutorial provides an end-to-end coding guide for running OpenAI's open-weight GPT-OSS models in Google Colab, with a focus on technical behavior, deployment requirements, and advanced inference workflows.

Key Points

  • Guides how to run OpenAI's open-weight GPT-OSS models in Google Colab.
  • Covers setting up dependencies for Transformers-based execution, checking GPU availability, and loading models with MXFP4 quantization and torch.bfloat16 activation.
  • Explores key capabilities such as structured generation, streaming, multi-turn conversation handling, tool execution patterns, and batch inference.
  • Highlights the trade-offs of open-weight models versus closed hosted APIs in terms of transparency, controllability, memory constraints, and local execution.
  • Treats GPT-OSS not merely as a chatbot, but as a technically inspectable open-weight LLM stack that can be configured, prompted, and scaled within reproducible workflows.
Notable Quotes & Details

Notable Data / Quotes

  • gpt-oss-20b (model name)
  • ~16GB VRAM (required VRAM)

Intended Audience

AI developers, machine learning engineers, researchers

Notes: A tutorial-format coding guide that includes real code snippets.

The Devil Is in Gradient Entanglement: Energy-Aware Gradient Coordinator for Robust Generalized Category Discovery

2026-04-18

Summary

A paper proposing an Energy-Aware Gradient Coordinator (EAGC) to address the gradient entanglement problem that arises in Generalized Category Discovery (GCD).

Key Points

  • Discovered that optimization interference in existing GCD methodologies hinders performance improvement.
  • Gradient entanglement weakens discriminability among known classes and causes representation space overlap with new classes.
  • EAGC consists of Anchor-based Gradient Alignment (AGA) and Energy-aware Elastic Projection (EEP).
  • AGA uses a reference model to fix the gradient direction of labeled samples, preserving the discriminative structure.
  • EEP projects unlabeled gradients onto the complementary subspace of the known-class subspace to reduce overlap.
  • Experiments demonstrate that EAGC improves the performance of existing methods and achieves state-of-the-art results.
Notable Quotes & Details

Intended Audience

AI researchers, machine learning developers

MixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLM Midtraining

2026-04-18

Summary

Introduces MixAtlas, an uncertainty-aware data mixture optimization method for multimodal LLM midtraining, which generates benchmark-targeted data recipes.

Key Points

  • Data mixture optimization for multimodal midtraining has primarily been conducted along a single dimension.
  • MixAtlas decomposes the training corpus along two axes: image concepts (10 visual domain clusters via CLIP embeddings) and task supervision (5 objective types including captioning, OCR, grounding, detection, and VQA).
  • Uses a small proxy model (Qwen2-0.5B) with a Gaussian Process surrogate model and GP-UCB acquisition to explore the mixture space.
  • On Qwen2-7B, the optimized mixture improves average performance by 8.5%–17.6% over the strongest baseline.
  • On Qwen2.5-7B, it achieves 1.0%–3.3% performance improvement and requires up to 2x fewer steps to reach the same training loss as the baseline.
  • Recipes found with the 0.5B proxy transfer to 7B-scale training across the Qwen model family.
Notable Quotes & Details

Notable Data / Quotes

  • 8.5%-17.6%
  • 1.0%-3.3%
  • 2 times fewer steps

Intended Audience

AI researchers, multimodal LLM developers

Portfolio Optimization Proxies under Label Scarcity and Regime Shifts via Bayesian and Deterministic Students under Semi-Supervised Sandwich Training

2026-04-18

Summary

A paper proposing a machine learning-based portfolio optimization framework for situations of data scarcity and market regime shifts, using labels generated by a Conditional Value at Risk (CVaR) optimizer to train neural network models.

Key Points

  • Proposes a machine learning-based portfolio optimization framework designed for low-data environments and regime uncertainty.
  • Builds a teacher-student training pipeline in which a Conditional Value at Risk (CVaR) optimizer generates supervised labels.
  • Trains Bayesian and deterministic neural network models using real and synthetically augmented data.
  • Generates synthetic data using factor-based models with t-copula residuals, enabling training beyond the limited 104 labeled observations.
  • Evaluates four student models in a structured experimental framework including controlled synthetic experiments, in-distribution real market evaluation, and cross-universe generalization.
  • In real market environments, models are deployed with a rolling evaluation protocol where they are periodically fine-tuned and reset to a base state.
  • Student models match or outperform the CVaR teacher in multiple settings, achieving improved robustness and reduced trading turnover during regime shifts.
Notable Quotes & Details

Notable Data / Quotes

  • 104 labeled observations

Intended Audience

Financial engineering researchers, quantitative investors, machine learning developers

Towards Verified and Targeted Explanations through Formal Methods

2026-04-18

Summary

A paper introducing the ViTaX (Verified and Targeted Explanations) framework, which leverages formal methods to provide trustworthy and targeted explanations for deep learning models.

Key Points

  • Emphasizes the need for interpretable and trustworthy explanations for deep neural networks deployed in safety-critical domains.
  • Points out limitations of existing XAI methodologies: heuristic attribution techniques lack mathematical guarantees about decision boundaries, and formal methods verify robustness but are not targeted.
  • ViTaX is a formal XAI framework that generates targeted counterfactual explanations with mathematical guarantees.
  • ViTaX identifies the minimal feature subset most sensitive to a specific y→t transition and formally guarantees via reachability analysis that perturbing these features by ε will not change the classification to t.
  • Formally guarantees how resilient a model is against user-identified alternatives through targeted ε-robustness.
  • Evaluations on MNIST, GTSRB, EMNIST, and TaxiNet demonstrate over 30% fidelity improvement with minimal explanation cardinality.
Notable Quotes & Details

Notable Data / Quotes

  • 30% fidelity improvement

Intended Audience

AI researchers, safety-critical systems developers, XAI researchers

Shapley Value-Guided Adaptive Ensemble Learning for Explainable Financial Fraud Detection with U.S. Regulatory Compliance Validation

2026-04-18

Summary

A study proposing the necessity of explainable AI models for regulatory compliance in US financial fraud detection, and demonstrating the effectiveness of a SHAP-based ensemble learning model.

Key Points

  • Faces regulatory compliance issues due to the inexplicability of AI-based fraud detection models.
  • Evaluates fidelity and stability of various explanation techniques (XGBoost + TreeExplainer shows high stability; LSTM + DeepExplainer shows low stability).
  • Proposes the SHAP-Guided Adaptive Ensemble (SGAE) model and achieves the highest AUC-ROC performance.
  • Evaluates three architectures — LSTM, Transformer, GNN-GraphSAGE — on a dataset of over 590,000 IEEE-CIS transactions.
  • All results are directly mapped to OCC, SR 11-7, and BSA-AML regulatory compliance requirements.
Notable Quotes & Details

Notable Data / Quotes

  • $32 billion each year
  • OCC Bulletin 2011-12
  • Federal Reserve SR 11-7
  • BSA-AML
  • W=0.9912
  • AUC-ROC 0.8837 (held-out); 0.9245 (cross-validation)
  • AUC-ROC 0.9248 and F1=0.6013
  • 590,540-transaction IEEE-CIS dataset

Intended Audience

AI researchers, financial fraud prevention specialists, regulatory compliance officers

oh-my-customcode — A tool that 'compiles' Claude Code agents instead of 'configuring' them

2026-04-18

Summary

Introduction of the `oh-my-customcode` tool, which introduces the concept of 'compilation' to solve the repetitive configuration problems that arise in Claude Code agent development.

Key Points

  • Automates the complex configuration process — including skills, YAML, and routing — when developing Claude Code agents.
  • Adopts the motto `Your AI Agent Stack. Compiled, Not Configured.` and separates reusable knowledge/workflows (skills) from agents.
  • A single `omcustom init` command provides 48 agents, 107 skills, 22 rules, and 39 guides.
  • The main conversation acts as a singleton orchestrator, with all tasks delegated to dedicated agents to prevent context mixing.
  • Model tiering is explicitly applied (opus for architecture/research, sonnet for implementation/agent creation, haiku for search/count verification).
  • Safety hooks (secret-filter, audit-log, etc.) operate in an advisory mode that only leaves warnings without blocking.
Notable Quotes & Details

Notable Data / Quotes

  • 48 agents / 107 skills / 22 rules / 39 guides
  • R010
  • opus
  • sonnet
  • haiku
  • reasoning-sandwich pattern
  • GitHub: https://github.com/baekenough/oh-my-customcode
  • npm: https://www.npmjs.com/package/oh-my-customcode

Intended Audience

Claude Code developers, AI agent system architects, productivity tool developers

Qwen3.5 model quantization — why do community versions underperform?

2026-04-18

Summary

Research analyzing the causes of performance degradation in community-version quantization of the Qwen3.5 model, and proposing mixed-bit quantization solutions that account for per-layer sensitivity.

Key Points

  • Tool call errors, meaningless outputs, and hallucination were observed in community-distributed MLX-format Qwen3.5 models.
  • Unsloth identified the causes and solutions through over 150 benchmark experiments.
  • Uniform quantization damages the sensitive `linear_attn.out_proj` layer without considering Qwen3.5's hybrid architecture (self-attention/GatedDeltaNet).
  • As a solution, proposes mixed-bit quantization based on per-layer sensitivity (3-bit for less sensitive MLP layers, 5-bit + AWQ for attention Q/K/V, bf16 for output layers).
  • Uses conversation, coding, and tool call examples as calibration data to reflect real-world importance.
  • Advantages: improved quality for tool calls, structured outputs, and code generation; equivalent performance to the same GGUF version and MLX.
  • Disadvantage: larger disk footprint than pure low-bit models due to retaining some layers in bf16.
Notable Quotes & Details

Notable Data / Quotes

  • MLX format Qwen3.5
  • Unsloth
  • over 150 benchmark experiments
  • 121 configuration comparisons
  • `linear_attn.out_proj` layer
  • Information loss sensitivity at 4-bit compression is approximately 120x higher than the output layer (lm_head)
  • MLP layers at 3-bit
  • Attention Q/K/V layers at 5-bit + AWQ
  • Output layer in bf16
  • Qwen3.6-35B-A3B
  • Claude Opus 4.7

Intended Audience

AI model developers, quantization researchers, MLX/LLM users, MLOps engineers

Notes: Contains praise for Unsloth founder Daniel Han.

Smol machines — sub-second cold starts, portable virtual machines

2026-04-18

Summary

Introduction of `smolvm`, a CLI-based virtual machine management tool supporting sub-second cold starts, elastic memory management, and single-file portability on macOS and Linux.

Key Points

  • `smolvm` is a CLI-based virtual machine management tool for running software in isolated environments.
  • Provides sub-second (under 1 second) cold starts, elastic memory management, and single-file portability for fast and lightweight VM execution.
  • VMs run as Linux kernel-based microVMs, packaged as `.smolmachine` files that can be re-run without dependencies.
  • Supports integration of development and security environments with hypervisor boundary isolation, SSH agent forwarding, and Smolfile-based environment declaration.
  • Supports booting OCI images without a Docker daemon, with boot times under 200ms and hardware-level isolation.
  • On macOS, runs an independent kernel on top of Hypervisor.framework; on Linux, runs on KVM.
Notable Quotes & Details

Notable Data / Quotes

  • `smolvm`
  • sub-second (under 1 second) cold start
  • boot time under 200ms
  • .smolmachine file
  • OCI image format
  • 4 vCPU, 8GiB RAM
  • Apache-2.0 license
  • @binsquare

Intended Audience

Developers, DevOps engineers, security professionals, virtualization technology users

Show GN: Make web Gemini look like VSCode — Gemini VSCode Theme Chrome Extension

2026-04-18

Summary

A Chrome extension has been developed that makes the web Gemini interface look like VS Code.

Key Points

  • Provides a VS Code-like UI including VS Code Dark+ theme, line numbers, activity bar, sidebar, status bar, and title bar.
  • Applies monospace fonts (JetBrains Mono, Noto Sans KR) and a terminal-style input box.
  • Turning on Python mode makes Gemini chat appear like Jupyter Notebook cells.
  • There are no additional VS Code features; it is simply a theme change.
Notable Quotes & Details

Intended Audience

Gemini users, developers, VS Code users

Show GN: Nilbox — Run OpenClaw without exposing your API token

2026-04-18

Summary

Nilbox is a tool that helps run OpenClaw safely without directly exposing API tokens to agents.

Key Points

  • Instead of passing real API tokens directly to agents, it replaces dummy tokens with real tokens at the network layer.
  • In the event of a token leak, all an attacker obtains is a meaningless string, enhancing security.
  • Supports macOS, Windows, and Linux, and provides a managed Linux runtime, a Store for one-click app installation, and full shell access.
Notable Quotes & Details

Intended Audience

AI agent developers, security-conscious users

ICML 2026 - Heavy score variance among various batches? [D]

2026-04-18

Summary

Questions and discussion about heavy score variance across batches in ICML 2026 paper reviews.

Key Points

  • Some batches have almost no papers above 3.5, while others record an average of 3.75, indicating large score differences.
  • Questions are raised about whether such score variance is due to domain differences, reviewer strictness, and whether ICML accounts for this.
Notable Quotes & Details

Notable Data / Quotes

  • 3.5 score
  • 3.75 average

Intended Audience

Machine learning researchers, conference reviewers, paper submitters

Zero-shot World Models Are Developmentally Efficient Learners [R]

2026-04-18

Summary

Zero-shot World Models (ZWM) achieve visual capabilities with far less data than human children, laying the groundwork for data-efficient AI systems.

Key Points

  • Current AI requires far more data than human children to achieve visual capabilities.
  • ZWM, trained on only a single child's visual experience, shows performance comparable to state-of-the-art models on a variety of visual-cognitive tasks.
  • Operates zero-shot without task-specific training, presenting a blueprint for developing data-efficient AI systems.
Notable Quotes & Details

Intended Audience

AI researchers, machine learning developers, AI ethics researchers

We're proud to open-source LIDARLearn [R] [D] [P]

2026-04-18

Summary

LIDARLearn, a unified PyTorch library for 3D point cloud deep learning, has been open-sourced.

Key Points

  • The first unified library supporting 56 configurations and cross-validation.
  • Easily runnable with a single YAML file, and automatically generates a LaTeX PDF report after training.
  • Includes benchmarks for various datasets such as ModelNet40, ShapeNet, and S3DIS.
  • Targets researchers in 3D point cloud learning, 3D computer vision, and remote sensing.
Notable Quotes & Details

Intended Audience

AI researchers, deep learning developers

easyaligner: Forced alignment with GPU acceleration and flexible text normalization (compatible with all w2v2 models on HF Hub) [P]

2026-04-18

Summary

easyaligner, a forced alignment library with GPU acceleration and flexible text normalization support, has been released.

Key Points

  • A high-performance, easy-to-use library designed for audio and text preprocessing.
  • Automatically detects relevant audio regions when transcripts do not cover all speech content.
  • Can process long audio and text segments without chunking.
  • Supports wav2vec2 models and can perform forced alignment of audio and text in various languages.
Notable Quotes & Details

Intended Audience

Machine learning developers, speech-to-text model researchers

Notes: Content incomplete

Gemma 4 actually running usable on an Android phone (not llama.cpp)

2026-04-18

Summary

Shares how to run the Gemma 4 LLM locally and smoothly on an Android phone using Google's LiteRT setup.

Key Points

  • Uses Google's LiteRT instead of llama.cpp to run Gemma 4 efficiently on Android.
  • Runs the LLM locally, automates its own apps via ADB, and works offline.
  • Provides detailed information and code to help users build their own local AI assistant.
Notable Quotes & Details

Intended Audience

AI developers, Android users, LLM enthusiasts

AI helped me build a custom PC and 4 apps in 6 months with zero coding experience

2026-04-18

Summary

A story about a user with no coding experience who, with the help of AI, assembled a custom PC and developed 4 apps.

Key Points

  • Successfully created a custom PC parts list using AI (ChatGPT and Claude).
  • AI was a huge help in developing 4 apps in 6 months with no coding experience.
  • Emphasizes that AI is not just a trend but the future, and strongly encourages learning AI.
Notable Quotes & Details

Intended Audience

General readers, AI beginners

I made a self healing PRD system for Claude code

2026-04-18

Summary

A self-healing PRD (Product Requirements Document) system was developed for Claude Code, which autonomously finds and resolves issues that arise during project development.

Key Points

  • Requests information needed for the PRD and reviews existing code to answer questions.
  • Splits plans into multiple files and begins the next step only after the previous step is complete.
  • Performs an independent review of the code via Codex after each step is completed.
  • When improving existing projects, continuously discovers and resolves new issues through Codex feedback.
  • The system finds and resolves issues autonomously as it scales the code.
Notable Quotes & Details

Intended Audience

Developers, AI system designers

Open-source list of GenAI-related incidents

2026-04-18

Summary

An open-source list of cases that highlight ethical issues with GenAI use is shared, sparking discussion about the use and limitations of LLMs.

Key Points

  • An open-source list collecting cases of ethical issues related to GenAI use.
  • Shared to encourage discussion about the use and limitations of LLMs.
Notable Quotes & Details

Intended Audience

AI researchers, developers, policy makers, general readers

Update on my February posts about replacing RAG retrieval with NL querying — some things I've learned from actually building it

2026-04-18

Summary

An update to a previous post about replacing RAG retrieval with natural language querying, sharing lessons learned from actually building it.

Key Points

  • Proposed the idea of an LLM storing a context window as a document store and querying it with natural language to replace embedding similarity-based search.
  • Pure semantic search can underperform not because of scale, but because queries and target content use different vocabulary, causing missed retrievals.
  • The solution is an 'index-first' strategy of narrowing candidates with a lightweight topic tag index before running natural language queries.
  • Claude tends to prefer internal reasoning and resist querying the memory store, requiring query requirements to be encoded in the system prompt.
  • If permanent state lives in the document store rather than the model, the interface LLM should be interchangeable.
Notable Quotes & Details

Intended Audience

LLM developers, AI researchers, RAG system designers

Notes: Content truncated.

RTX 5070 Ti + 9800X3D running Qwen3.6-35B-A3B at 79 t/s with 128K context, the --n-cpu-moe flag is the most important part.

2026-04-18

Summary

Covers how to run the Qwen3.6-35B-A3B model at 79 t/s with 128K context in an RTX 5070 Ti and Ryzen 9800X3D environment, highlighting the importance of the `--n-cpu-moe` flag.

Key Points

  • Points out that a typical `--cpu-moe` setting causes a 54% speed loss on a 16GB GPU.
  • The `--n-cpu-moe N` flag keeps the experts of the first N layers on the CPU and places the rest on the GPU, enabling efficient VRAM usage.
  • With N=20, achieved a 54% improvement in generation and prompt speed.
  • 128K context can be scaled almost for free thanks to the `-np 1` setting.
  • Used Claude Opus 4.7 to autonomously perform setup construction, benchmark execution, and tuning iterations.
Notable Quotes & Details

Notable Data / Quotes

  • RTX 5070 Ti (16GB GDDR7)
  • Ryzen 9800X3D
  • 32GB DDR5 RAM
  • Qwen3.6-35B-A3B model (22.1 GB)
  • 79 t/s (token per second)
  • 128K context
  • 54% generation speed, +54% prompt speed
  • `--cpu-moe` (baseline): 51.2 Gen t/s, 87.9 Prompt t/s, 3.5 GB VRAM
  • `--n-cpu-moe 20`: 78.7 Gen t/s, 100.6 Prompt t/s, 12.7 GB VRAM
  • `--n-cpu-moe 20 + -np 1 + 128K ctx`: 79.3 Gen t/s, 135.8 Prompt t/s, 13.2 GB VRAM

Intended Audience

LLM developers, hardware enthusiasts, high-performance computing users

Notes: Content truncated.

Cloudflare open-sources lossless LLM compression tool

2026-04-18

Summary

Cloudflare open-sources Unweight, a lossless compression system that reduces LLM size by 15-22%, saving VRAM usage.

Key Points

  • Cloudflare releases lossless LLM compression system 'Unweight'.
  • Reduces LLM size by 15-22%, saving VRAM usage.
  • Saves approximately 3GB of VRAM on Meta's Llama-3.1-8B model with Nvidia H100 GPUs.
  • Focuses on MLP weight compression, with plans to extend to attention weights in the future.
  • GPU kernels open-sourced on GitHub, with a technical paper published.
Notable Quotes & Details

Notable Data / Quotes

  • 15–22%
  • 3 GB
  • Llama-3.1-8B

Intended Audience

AI developers, LLM researchers, systems engineers

Qwen 3.6 35B A3B Q4_K_M quant evaluation

2026-04-18

Summary

The Q4_K_M quantized version of the Qwen 3.6 35B A3B model was evaluated on code generation, commonsense reasoning, and function call benchmarks in a CPU environment.

Key Points

  • Evaluation of the Q4_K_M quantized version of the Qwen 3.6 35B (3B active MoE) model.
  • Run with llama-cpp-python in a CPU environment (32 vCPU, 125GB RAM).
  • Recorded HumanEval (code generation) 47.56%, HellaSwag (commonsense reasoning) 74.30%, BFCL (function calling) 46.00%.
  • Best performance in commonsense reasoning, with a speed of 22 tokens/sec on CPU.
  • Evaluation conducted via Neo AI Engineer.
Notable Quotes & Details

Notable Data / Quotes

  • 35B
  • 3B
  • 47.56%
  • 74.30%
  • 46.00%
  • 22 tokens/sec

Intended Audience

AI researchers, LLM developers, quantized model users

Should you shut off thinking when you are coding on say Qwen3.6 35B

2026-04-18

Summary

Discussion about whether to disable the AI's 'thinking' feature when coding with models like Qwen3.6 35B, and how to configure this in LM Studio.

Key Points

  • Claims that an AI model's 'thinking' feature can slow down the system.
  • Views the 'thinking' feature as analogous to a 'to-do list' in Claude Code or Codex.
  • Opinion that having AI create a 'to-do list' without fully relying on the model may be better.
  • Difficulty finding how to disable the 'thinking' feature for the Qwen3.6 35B model in LM Studio.
Notable Quotes & Details

Notable Data / Quotes

  • "Qwen3.6 35B"

Intended Audience

LLM users, AI developers, LM Studio users

LM Studio CPU thread pool size vs. tk/s with some MoE layers offloaded to CPU

2026-04-18

Summary

Covers the relationship between CPU thread pool size and tokens per second (tk/s) when some MoE layers are offloaded to CPU in LM Studio.

Key Points

  • Related to CPU offloading of MoE layers in the LM Studio environment.
  • Analysis of the impact of CPU thread pool size on token processing speed (tk/s).
  • Only a link is provided without specific details.
Notable Quotes & Details

Intended Audience

LLM developers, LM Studio users, systems engineers

Notes: Content incomplete

AWS Announces General Availability of DevOps Agent for Automated Incident Investigation

2026-04-18

Summary

AWS has officially launched its generative AI-powered DevOps Agent, enabling developers and operators to troubleshoot, analyze deployments, and automate operational tasks in AWS environments.

Key Points

  • DevOps Agent was previewed at re:Invent 2025 and is built on Amazon Bedrock AgentCore.
  • The agent learns application relationships and integrates with observability tools, runbooks, code repositories, and CI/CD pipelines to analyze incidents.
  • Correlates telemetry, code, and deployment data to autonomously triage issues, accelerate resolution, and recommends improvements by identifying historical incident patterns to prevent future outages.
  • With general availability, new capabilities were added: the ability to investigate applications in Azure and on-premises environments, custom agent skill support to extend functionality, and custom charts and reports.
  • DevOps Agent is not a passive Q&A tool but an autonomous team member that immediately begins investigations when event sources such as CloudWatch alarms and PagerDuty alerts are triggered.
Notable Quotes & Details

Notable Data / Quotes

  • re:Invent 2025
  • Madhu Balaji
  • Janardhan Molumuri
  • Bill Fine
  • Joe Alioto
  • Tipu Qureshi

Intended Audience

DevOps engineers, SREs, cloud operations managers

[Webinar] Eliminate Ghost Identities Before They Expose Your Enterprise Data

2026-04-18

Summary

A webinar covering how to find and eliminate 'Ghost Identities' before they expose an organization's enterprise data.

Key Points

  • 68% of cloud breaches in 2024 were caused by unmanaged non-human identities such as service accounts and forgotten API keys.
  • There are 40–50 automated credentials (service accounts, API tokens, AI agent connections, OAuth grants) per employee, and most remain active even after a project ends or an employee leaves.
  • AI agents and automated workflows are multiplying these credentials at a pace that security teams cannot track manually.
  • Many of these credentials have admin-level access they don't need, and a single compromised token can enable lateral movement across an entire environment.
  • This session covers: how to run a full discovery scan of all non-human identities in an environment, a framework for right-sizing permissions across service accounts and AI integrations, and automated lifecycle policies to ensure expired credentials are revoked before attackers find them.
Notable Quotes & Details

Notable Data / Quotes

  • 68% of cloud breaches in 2024
  • average dwell time of over 200 days

Intended Audience

Security administrators, IT managers, CISOs

Mirai Variant Nexcorium Exploits CVE-2024-3721 to Hijack TBK DVRs for DDoS Botnet

2026-04-18

Summary

A Mirai variant called Nexcorium is exploiting the CVE-2024-3721 vulnerability to hijack TBK DVRs for a DDoS botnet.

Key Points

  • Attackers are exploiting security vulnerabilities in TBK DVRs and end-of-life (EoL) TP-Link Wi-Fi routers to deploy Mirai botnet variants.
  • CVE-2024-3721 is a medium-severity command injection vulnerability affecting TBK DVR-4104 and DVR-4216 digital video recording devices.
  • IoT devices are a prime target for large-scale attacks due to their widespread use, lack of patches, and weak security configurations.
  • Nexcorium has an architecture similar to Mirai variants, including XOR-encoded configuration table initialization, a Watchdog module, and DDoS attack modules.
  • The malware also includes an exploit for CVE-2017-17215 targeting Huawei HG532 devices, and incorporates hardcoded lists of usernames and passwords used for brute-force attacks.
Notable Quotes & Details

Notable Data / Quotes

  • CVE-2024-3721 (CVSS score: 6.3)
  • CVE-2017-17215
  • September 2025
  • Vincent Li

Intended Audience

Network administrators, security analysts, IoT device users

Jensen Huang delivers a direct blow to US tech regulations: 'defeatist and lunacy'

2026-04-18

Summary

Nvidia CEO Jensen Huang harshly criticized US technology regulatory policy as 'defeatist and lunacy', expressing concern about China's acceleration in building its own AI stack.

Key Points

  • Huang appeared on the Dwarkesh Patel podcast and firmly rebutted scenarios in which either Nvidia keeps selling chips to China and China acquires the technology to overtake the US, or regulations cause Nvidia to lose the Chinese market.
  • He pointed out that abandoning a market is true defeatism, and emphasized that Nvidia is not merely a hardware manufacturer but a company with a complex software ecosystem (CUDA).
  • Criticized US export controls for actually accelerating China's development of an independent AI stack centered on Huawei chips.
  • Cited the fact that nearly half of the world's AI researchers are in China, warning that if global AI models run well on non-US technology stacks, it would be a horrible outcome for the United States.
  • Strongly rebutted the claims of Anthropic CEO Dario Amodei and others who liken AI to nuclear weapons and call for tighter regulations, calling such views 'lunacy'.
Notable Quotes & Details

Notable Data / Quotes

  • 2026-04-16 (local time)
  • Dwarkesh Patel podcast
  • 400,000 views
  • 2,200+ comments
  • "You are not talking to somebody that woke up a loser. That loser attitude and loser premise makes no sense to me. We are not a car."

Intended Audience

AI industry stakeholders, policy makers, tech news readers

Mythos alert spreading globally — UK and India demand early access to the model

2026-04-18

Summary

Anthropic's next-generation AI model 'Claude Mythos' is becoming a target of vigilance from global financial institutions and regulators due to its potential cybersecurity threat, with the UK, India, and others demanding early access.

Key Points

  • Anthropic's Claude Mythos is assessed as having the potential to neutralize existing cybersecurity frameworks.
  • Thousands of 'zero-day' vulnerabilities were discovered during model testing, presenting a dual-use nature with both security enhancement and exploitation potential.
  • Policy authorities and financial institutions around the world are rapidly responding to the risks of Mythos, with the UK, India, and others demanding early access.
  • Regulation and response frameworks are lagging behind the pace of technological advancement, and international cooperation frameworks are still in their early stages.
  • Anthropic is aware of the model's risks and is minimizing the potential for misuse through testing in restricted environments and safety measures.
Notable Quotes & Details

Notable Data / Quotes

  • Bloomberg, April 16 (local time)
  • April 17
  • Andrew Bailey, Governor of the Bank of England
  • Christine Lagarde, President of the European Central Bank
  • "If it falls into the wrong hands, it could have serious consequences"

Intended Audience

Cybersecurity professionals, financial institution officials, policy makers, AI developers

OpenAI, three executives resign including VP of Science team

2026-04-18

Summary

Three key executives at OpenAI have resigned, including VP of Science Kevin Weil, and this appears to have occurred as the company restructures around B2B business and a 'super app' focus.

Key Points

  • OpenAI VP of Science Kevin Weil, Sora team leader Bill Peebles, and B2B Applications CTO Srinivas Narayanan resigned on April 17 (local time).
  • Their resignations are related to OpenAI's moves to focus on B2B business and consolidate the organization around a 'super app'.
  • VP Kevin Weil stated he is leaving because 'the science team is being distributed across other departments'.
  • OpenAI plans to strengthen collaboration with other teams through the decentralization of the science team and focus on developing an 'AI researcher'.
  • There have been a series of recent executive reshuffles, including a leave of absence by OpenAI Applications CEO Fifi Simo and role changes for other executives.
Notable Quotes & Details

Notable Data / Quotes

  • April 17 (local time)
  • June 2024
  • Today is my last day at OpenAI, as OpenAI for Science is being decentralized into other research teams. It's been a mind-expanding two years, from Chief Product Officer to joining the research team and starting OpenAI for Science. Accelerating science will be one of the most…

Intended Audience

AI industry stakeholders, investors, OpenAI users

Anthropic unveils all-in-one design tool 'Claude Design', threatening Adobe and Canva

2026-04-18

Summary

Anthropic has unveiled 'Claude Design', a conversational AI-powered all-in-one design tool, threatening the existing design software market dominated by Adobe and Canva.

Key Points

  • Anthropic unveiled 'Claude Design' on April 17 (local time), a visual design and prototyping tool powered by conversational AI.
  • Built on 'Claude Opus 4.7', it generates design drafts through natural language descriptions and refines outputs through conversation and editing.
  • Provides a 'design system' feature that reads a company's codebase and design files to automatically reflect colors, typography, and UI components.
  • Has a 'handoff' structure that passes completed designs to the AI coding tool 'Claude Code' for implementation, enabling one-stop processing from idea to deployment.
  • Designed so that non-experts can create high-quality designs, and is expected to have a significant impact on the existing design software market.
Notable Quotes & Details

Notable Data / Quotes

  • April 17 (local time)
  • Claude Opus 4.7

Intended Audience

Designers, product managers, marketers, founders, AI and design software industry stakeholders

[April 17] The warning of 'cognitive surrender': 'Capable AI stops humans from thinking'

2026-04-18

Summary

Researchers at the Wharton School of the University of Pennsylvania are warning, through the concept of 'cognitive surrender', that capable AI stops humans from making critical thinking efforts and leads them to uncritically accept AI outputs.

Key Points

  • 'Cognitive surrender' refers to the phenomenon in which humans stop critically thinking about AI answers and uncritically accept AI outputs.
  • The more capable AI becomes, the more humans tend to rate the authority of AI algorithms higher than their own knowledge and relinquish intellectual initiative.
  • When AI presents wrong answers, groups using AI chose the wrong answer at a higher rate and spent less time analyzing than when solving problems on their own.
  • When AI's reliability exceeds a certain level, humans tend to stop critical thinking, operating in the structure of 'AI accuracy rises → trust increases → cost of human verification increases → verification abandoned'.
  • Warns that AI could become mental infrastructure, and that 'not an inability to think, but a state of feeling no need to think' could be the real problem of the AI era.
Notable Quotes & Details

Notable Data / Quotes

  • 'Cognitive Surrender'
  • Wharton School, University of Pennsylvania
  • last February
  • 'The Speed of Thought: How AI Is Changing the World'

Intended Audience

AI researchers, general readers, policy makers, cognitive science researchers

Security TF member Yoon Du-sik: 'Korea is not in a zero-day defense environment'

2026-04-18

Summary

Yoon Du-sik, a member of the National AI Strategy Committee's Security TF, delivered a presentation on 'National AI Security Strategy and Basic Direction in the AI Era', explaining Korea's deficiencies in the zero-day defense environment and key AI security challenges.

Key Points

  • Korea must establish security as the foundation of an AI-based society in order to achieve its goal of becoming one of the world's top 3 AI nations.
  • Domestic breach incident reports are increasing by more than 15% per year, and the zero-day defense environment is inadequate.
  • The National AI Strategy Committee's AI Action Plan includes three pillars: fostering an AI innovation ecosystem, a nationwide AI-based transformation, and contributing to a global AI-based society.
  • Key security challenges include: activating the public and private AI security ecosystem (ISMS-P reform, CVD/VDP, nurturing white-hat hackers and the security industry), building an AI-based cybersecurity framework (K-Cyber Security LLM, AI-ISAC, security internalization), and strengthening AI security threat response and cooperation (cyber security platform, CBRN preparedness, technology leak prevention).
  • In particular, the introduction of a CVD/VDP (Coordinated Vulnerability Disclosure) operation framework is being promoted, which is a process for systematically reporting security vulnerabilities and disclosing them after remediation.
Notable Quotes & Details

Notable Data / Quotes

  • "Domestic breach incident reports to KISA (Korea Internet & Security Agency) are increasing by more than 15% annually."
  • "We plan to improve ISMS-P (Information Security and Personal Information Protection Management System certification) by incorporating an 'attack surface management inspection' method, enabling companies to identify and manage vulnerabilities in their IT assets, with a target implementation in Q2 of this year."

Intended Audience

Information security professionals, AI policy makers, corporate security officers

[Ahn Kwang-seop's AI Synthesis] The dizzy world of vibe marketing — 'vibe' is nothing but a pretense

2026-04-18

Summary

Writer Ahn Kwang-seop criticizes the 'vibe marketing' phenomenon, explaining the essence of technology and the knowledge moat that AI is demolishing.

Key Points

  • Presents a critical view that 'vibe' is used as a marketing term in Korea to package the appearance of being ahead in the AI era, but in reality it is nothing but a pretense.
  • Even Andrej Karpathy, who coined 'vibe coding', began distancing himself from the term a year later, saying 'agentic engineering' would be more appropriate.
  • At ICML, 497 papers were rejected for unauthorized use of LLMs — the 'vibe paper' phenomenon — showing the reality that even papers are being replaced by mere pretense.
  • The essence of technology is 'eliminating the moat', and AI is rapidly demolishing the 'knowledge moat' that was once protected by accumulated knowledge and expertise.
  • Points out that expressions like AX (AI Transformation) are being used as marketing tools that merely replace names without substantive change.
Notable Quotes & Details

Notable Data / Quotes

  • Andrej Karpathy first used the term 'vibe coding' in February 2025 and began distancing himself from it in February 2026.
  • In March 2026, 497 papers (approximately 2% of total submissions) were rejected from ICML for unauthorized use of LLMs.

Intended Audience

AI industry professionals, marketers, technology trend analysts, general readers

Notes: Critical commentary

PreviousDaily Briefing
NextDaily Briefing
Jooojub
System S/W engineer
Explore Tags
Series
    Recent Post
    © 2026. jooojub. All right reserved.